Rahul G. Krishnan

Assistant Professor, Computer Science & Laboratory Medicine and Pathobiology
Canada CIFAR AI Chair at the Vector Institute
Tier II Canada Research Chair in Computational Medicine

Research Direction

Data collected from natural phenomena characterize computation occurring within complex processes governed by known and unknown laws. Neural networks are powerful tools that can learn to compress computation happening in nature. My group advances fundamental research in machine learning to understand, identify, and create controllable and reliable artificial intelligence systems.

Research Threads

• Causal machine learning. Natural phenomena follow laws, physics among them, that explain cause and effect. Humans generalize remarkably rapidly because we can manipulate or intuit those laws. Teaching neural networks how to do the same is critical to abduction and generalization, yet traditional tools for causality have resisted the bitter lesson. Our group has built and scaled the first causal foundation models (e.g. CausalPFN, IV-ICL and SurvivalPFN) which enable zero-shot prediction of causal effects, bounds on causal effects, and survival outcomes via in-context learning. [1] [2] [3] [4] [5]
• Generative models. Generative models compress the computation in nature directly. The more efficiently, in terms of FLOPs, we can compress that computation, the faster we can scale over data and parameters. We are interested in understanding how to minimize the distance between energy (via compute) and bits using generative models like MDM-Prime-v2 Deep Markov Models [1] [2] [3] [4]
• Large language models. Large language models compress computation expressed in natural language text. By prompting we can index into this computation. But to use LLMs reliably in health and biology, we need the indexed computation to mimic human intent. We also need new ways for LLMs to quickly acquire domain knowledge and ensure their outputs can be trusted in high-stakes settings. We develop methods like Contextual Fine-Tuning that teach LLMs how to learn quickly from new corpora, tools like AutoElicit to help experimentalists use LLMs as proxies for expert opinions, agentic systems AMG-RAG for question answering when knowledge evolves, and study causal methods to the development and evaluation of language models. [1] [2]
• Decision making in healthcare and biology. Solving decision-making problems in health at scale requires tools that can extract signal from large images, (HIPT), from long time series (SurF) and from tabular data (InterpreTabNet), and composing them correctly (CoMET). Deployment in medicine also requires creating tools to detect model failures like Detectron, and D3M. [1] [2] [3] [4] [5]

Students

PhD Students

• Vahid Balazadeh-Meresht
• Ethan Choi
• Viet Nguyen
• Lance Chao
• Yeongbin Seo
• Michael Cooper (co-sup w/ Mike Brudno)
• Jerry Ji (co-sup w/ Anna Goldenberg)
• Yujia Ma (co-sup w/ Chris Maddison)
• Steven Palayew (co-sup w/ Mike Wainberg)
• Mohammad Adnan (PhD student at the University of Calgary sup. w/ Yani Ioannou)

Postdoctoral Fellows

• Maria Diaz de Leon Derby
• Qiuhao Zeng
• David Pellow (Schmidt Fellow with Mike Brudno)

MSc Students

• Aditya Mehrotra

Links

• Bio
• Joining the Lab
• Invited Talks & Media
• Research Grants & Awards
• Teaching
• Writing
• Alumni

Selected Publications

A selected list of representative papers is available below. For a full list, see my Google Scholar profile.